The Promise and Limits of Machine Learning in Football Attacking Analysis
A review of how machine learning has been applied to analyze attacking performance, identify key indicators, and support tactical decision-making in professional football.
The following summary critically reviews the research paper titled “Machine Learning in Men’s Professional Football: Current Applications and Future Directions for Improving Attacking Play” by Mat Herold, Floris Goes, Stephan Nopp, Pascal Bauer, Chris Thompson, and Tim Meyer. All data, figures, and analysis presented here are drawn from their original work; I do not claim any authorship or ownership of the content. This summary has been written to provide a concise and technically informed synthesis of the paper’s findings, methodologies, and implications, while maintaining fidelity to the authors’ intellectual contributions.
Introduction
The paper opens by situating machine learning within the long-standing search for key performance indicators related to attacking success in elite football. Traditionally, match analysis has relied on notational approaches that summarize discrete events such as passes, shots, or entries into the penalty area. Prior research consistently showed that successful teams create more goal-scoring opportunities, penetrate defensive structures more effectively, and gain frequent access to the penalty box. However, the authors argue that these approaches are limited because they are time-consuming and largely ignore the interactional and contextual dynamics that continuously evolve during a match.
Recent technological advances, particularly in automatic tracking systems, video-based motion analysis, and GPS technologies, have expanded the analytical possibilities available to practitioners. These developments allow analysts to move beyond static event counts toward dynamic representations of team behaviour, including spatial dominance, coordination, and tempo. Within this context, machine learning is introduced as an emerging methodological framework capable of detecting complex patterns in large, high-dimensional datasets, especially positional tracking data.
The authors carefully distinguish between supervised and unsupervised learning. In supervised learning, labelled examples are used to approximate a function of the form “y = f(x)”, where y represents an outcome such as a pass type or match phase, and x denotes a set of features. Unsupervised learning, by contrast, seeks to uncover latent structures in unlabelled data, for example identifying player roles or team coordination patterns without predefined categories. While both approaches have shown promise, the authors stress that relatively little is known about their practical usefulness for improving attacking play. Consequently, the stated aim of the review is to provide a critical appraisal of existing machine learning research in football, focusing specifically on attacking performance, and to identify current challenges and future research directions.
Methods
The review adopts a descriptive approach, surveying literature published between 1996 and 2018. Searches were conducted across major scientific databases using combinations of terms related to football, machine learning, and performance analysis. Inclusion criteria were deliberately narrow: studies had to involve professional adult male footballers, focus on technical or tactical evaluation, and be written in English. Studies involving youth players, women’s football, futsal, set plays, or small-sided games were excluded.
The authors justify this focus by noting that match behaviour differs substantially from training contexts, small-sided games, and youth competitions, and that most machine learning research in football has been conducted using full-match data. Each included study was assessed according to its learning paradigm (supervised or unsupervised), specific algorithms used, data source (event data versus tracking data), and the degree to which methodological details were reported sufficiently to allow reproduction. This structured evaluation underpins the comparative analysis presented in the subsequent sections.
Machine Learning in Football
The core of the review is divided into two main strands: studies using event data and studies using tracking data. This distinction reflects both the historical development of football analytics and the differing analytical possibilities afforded by each data type.
Machine Learning Models in Football Using Event Data
Event data have long been the dominant source for performance analysis, capturing discrete actions such as passes, shots, tackles, and fouls. Early machine learning applications in this domain focused on identifying tactical patterns or predicting match outcomes. Unsupervised approaches, such as clustering of pass sequences, demonstrated that certain attacking patterns, including “side-attacks” or “zig-zag pass transactions”, could be automatically identified from temporal event data. However, these studies often relied on very small datasets and lacked proper validation, raising concerns about their robustness.
Later work using much larger datasets sought to link technical performance indicators to team success. Logistic regression and classification models applied to millions of events across multiple European leagues showed that successful teams typically produced more passes and shots while committing fewer fouls and defensive actions. Although these findings aligned with existing football knowledge, the authors note that such models still struggled to capture psychological and contextual factors, and that purely event-based representations cannot fully describe spatio-temporal dominance or interaction effects between teams.
A recurring theme in this section is the importance of spatial context. Studies examining passes into critical pitch zones, such as “Zone 14”, confirmed earlier notational findings that possession in central attacking areas correlates with shooting opportunities. Yet the authors caution that zone-based approaches are coarse abstractions. As they point out, “achieving possession in a specific zone does not guarantee or provide information about whether a team is penetrating the opponent’s defence”, since defensive organisation and situational context are ignored.
Event data have also been used for player recruitment and evaluation. Machine learning models identified performance profiles associated with progression to higher leagues, highlighting indicators such as pass accuracy and involvement in dangerous areas. Similarly, expected goals models represented a major step forward by quantifying shot quality on a probabilistic scale. Nevertheless, the authors argue that many xG models “did not acknowledge or capture opponent positioning” and therefore lacked tactical context, limiting their direct applicability for coaching decisions.
Overall, the review concludes that event-data-based machine learning studies have been valuable for identifying correlations between performance indicators and success, but they often rely on relatively simple models and struggle to represent the interactive and dynamic nature of attacking play.
Machine Learning Models Using Tracking Data
The second strand of the review focuses on studies using tracking data, which provide continuous information about the positions and movements of all players and the ball. The authors emphasize that tracking data enable more process-oriented analyses, allowing researchers to study how attacks develop over time and space.
Pass pattern recognition and classification
A substantial body of work has examined pass pattern recognition and pass quality classification. Early simulation-based models evaluated players’ ability to execute, receive, and refrain from passes, but required expert-labelled data to be practically useful. Subsequent supervised learning approaches classified passes as “Good”, “OK”, or “Bad” with high reported accuracy. However, the authors critically note that inter-observer agreement among human experts was only moderate, and that high classification accuracy does not necessarily imply meaningful tactical insight. As they observe, false positives and false negatives remained common, limiting the reliability of these classifications for coaching feedback.
More advanced models attempted to evaluate passes by their tactical effect rather than their immediate success. For example, data-driven measures of defensive disruption quantified how a pass altered the opponent’s defensive organisation over subsequent seconds. One key finding was that “not every pass can be an assist”, highlighting the value of sideways and backward passes in destabilising defensive structures. This represented an important conceptual shift away from purely forward-oriented notions of attacking effectiveness.
Team passing behaviour
Other studies analysed team-level passing behaviour, entropy, and build-up strategies. High-performing teams were found to exhibit greater unpredictability in their passing patterns, particularly near the penalty area. While such findings offered new descriptive insights, the authors note that they did not fully explain which off-ball movements or tactical principles generated this unpredictability.
Team behaviour related to time, space and goal-scoring
Tracking data have also been used to study the creation of goal-scoring opportunities. Models incorporating spatio-temporal features showed that expected goals depend not only on shot location but also on defender proximity, speed of play, and player interactions in the seconds preceding a shot. The authors quote one illustrative example where a specific attacking pattern “results in a chance of 70.59%”, but they caution that isolated probabilities do not explain how such situations arise or how they depend on player-specific or opponent-specific factors.
A particularly influential contribution reviewed is the concept of off-ball scoring opportunity, which estimates the likelihood that a player without the ball will score given the current game state. This approach highlighted players who consistently find dangerous spaces but are not rewarded with passes, offering potential applications in scouting and tactical preparation. Nonetheless, the authors point out limitations, including the absence of defensive pressure modelling and individual skill differentiation.
Defence and pressing tactics related to attacking play
Finally, the review discusses studies on pressing, counter-attacks, and defensive transitions. Machine learning models were able to automatically identify counter-attacks and quantify their offensive threat in short time windows after ball recovery. These approaches demonstrated the potential to exceed traditional notational analyses by capturing rapid transitions and spatial imbalances, although key questions remain about which specific parameters drive counter-attacking success.
Current Challenges and Future Directions of Machine Learning in Football
Current Challenges of Machine Learning in Football
In this section, the authors synthesise the limitations observed across the reviewed literature. A central challenge concerns data quality and validity, particularly for tracking systems that are prone to errors during fast, congested phases of play. The authors advise caution when comparing results across different tracking providers.
Another major issue is interpretability. Many machine learning models, especially complex neural networks, function as “black boxes”, making it difficult to explain why a particular pass or movement is rated highly. This lack of transparency conflicts with the preferences of coaches, who tend to favour simple, actionable insights and “quick snapshots” of performance. As the authors note, “performance analysis research including substantial and complex statistics and mathematical equations are not priorities for coaches, nor have they been successfully integrated into coaching”.
The review also highlights a disconnect between computer science research and applied football practice. Many studies focus on prediction tasks, such as match outcomes, that offer limited guidance for tactical improvement. Furthermore, the field lacks a cumulative learning curve, with researchers often prioritising novel methods over building on existing models. Early studies frequently suffered from small sample sizes and inadequate validation, although methodological standards have improved in more recent work.
Crucially, the authors argue that machine learning must aim to provide insights beyond what skilled human observers can already perceive. They cite examples, such as visual exploration behaviour before receiving the ball, where observational analysis still outperforms automated methods. Bridging this gap requires closer integration of football theory, contextual knowledge, and machine learning techniques.
Future Directions of Machine Learning in Football
Looking forward, the authors call for larger datasets, clearer methodological reporting, and consistent use of training and testing splits to ensure generalisability. Sharing code and workflows is encouraged to facilitate replication and cumulative progress.
Substantively, future research should focus on questions that are directly relevant to attacking play, such as how off-ball movement shapes passing decisions, how different defensive schemes can be penetrated, and how teams adapt their tempo and structure across match phases. The authors emphasise that progress will depend not only on more sophisticated algorithms but also on improved communication and collaboration between computer scientists, sports scientists, analysts, and coaches.
The overarching message is that machine learning holds significant promise for enriching tactical understanding in football, but only if it is grounded in football-specific theory and translated into forms that practitioners can meaningfully apply.
Conclusion
The authors conclude that, although machine learning offers powerful tools to analyse complex interactions in football, most existing work remains descriptive, overly technical, and poorly integrated into coaching practice. Much of the research has prioritised predictive accuracy and methodological novelty over practical relevance.
They argue that the real potential of machine learning lies in supporting tactical understanding, particularly when combined with the expertise of coaches and analysts. However, this requires simpler, interpretable outputs and stronger collaboration between computer scientists and football practitioners to ensure that analytical insights can be meaningfully applied on the pitch
Learn More
Here you have a direct link to get the free xG Mini-Guide, suited for those who want to fully understand the model in 5 minutes.
References
Herold, M., Goes, F., Nopp, S., Bauer, P., Thompson, C., & Meyer, T. (2019). Machine learning in men’s professional football: Current applications and future directions for improving attacking play. International Journal of Sports Science & Coaching, 14(6), 798-817. https://doi.org/10.1177/1747954119879350
Brilliant analysis.